Research On Geolocation Prediction Based On Social Networks

The continuous development of big data application in the present day has brought great convenience to our daily life.In return,it becomes especially critical to provide the user's location information which can be used to personalize their services.The rising of social media and the expansion of audiences have generated increasing amount of data,which contains huge potential location information of users.How to get the user's general location from such data has become a hot research topic.Research efforts have been devoted to analyzing and studying the complex data such as social network text and images from different angles.Many models have achieved good prediction results and laid a solid foundation for the geographical prediction task based on social network.This paper focuses on the geolocation technology based on social network,and puts forward a solution for the preprocessing of social network data,the key problems in positioning prediction,and the construction and deployment of the system.Most of the existing geolocation methods based on social network lack the scope of the mining of text characteristics,and the degree of integration of text characteristics and network features is low.Based on the work experience of previous people,this paper proposes a geolocation method based on multiview social network,the main work and innovation points are as follows:(1)The text information user published has been preprocessed from multiple angles with multi-level data cleaning.Inspired by multi-view learning,we extract its features from multiple views,including word view,sentence view and theme view,and text view features are made up of them.(2)In model construction,the emerging graph data processing technology—graph convolution neural network and graph attention neural network—has been considered.Based on the GCN and GAT networks,the text view features and the network view features built by friendship diagram can be fully fused.Experimental results show that the performance of our model is competitive to state-of-the-art best model.(3)the trained model is integrated into the web-based geographical prediction system.The system is designed using the style of RESTful.In particular,it can realize the separation of the frontend and backend in the development stage.In the application stage,it can provide users with APIs to fully meet users' requirements.We also conducted stress test of the system to provide users with trusted and reliable prediction service.